Ann Oper Res (2007) 153: 257–296 DOI 10.1007/s10479-007-0172-6
Complexity and algorithms for nonlinear optimization problems
Dorit S. Hochbaum
Published online: 3 May 2007 © Springer Science+Business Media, LLC 2007
Abstract Nonlinear optimization algorithms are rarely discussed from a complexity point of view. Even the concept of solving nonlinear problems on digital computers is not well defined. The focus here is on a complexity approach for designing and analyzing algorithms for nonlinear optimization problems providing optimal solutions with prespecified accuracy in the solution space. We delineate the complexity status of convex problems over network constraints, dual of flow constraints, dual of multi-commodity, constraints defined by a sub- modular rank function (a generalized allocation problem), tree networks, diagonal dominant matrices, and nonlinear knapsack problem’s constraint. All these problems, except for the latter in integers, have polynomial time algorithms which may be viewed within a unifying framework of a proximity-scaling technique or a threshold technique. The complexity of many of these algorithms is furthermore best possible in that it matches lower bounds on the complexity of the respective problems. In general nonseparable optimization problems are shown to be considerably more dif- ficult than separable problems. We compare the complexity of continuous versus discrete nonlinear problems and list some major open problems in the area of nonlinear optimiza- tion.
Keywords Nonlinear optimization · Convex network flow · Strongly polynomial algorithms · Lower bounds on complexity
1 Introduction
Nonlinear optimization problems are considered to be harder than linear problems. This is the chief reason why approximate linear models are frequently used even if the circum- stances justify a nonlinear objective. A typical approach is to replace an objective function
An earlier version of this paper appeared in 4OR, 3:3, 171–216, 2005. D.S. Hochbaum () Department of Industrial Engineering and Operations Research and Walter A. Haas School of Business, University of California, Berkeley, USA e-mail: [email protected] 258 Ann Oper Res (2007) 153: 257–296 that is nonlinear by a piecewise linear function. This approach may adversely affect the algo- rithm’s complexity as often the number of pieces is very large and integer variables may have to be introduced. For problems with nonlinear objective function the leading methodology consists of iterative and numerical algorithms in which complexity analysis is substituted with a convergence rate proof. In order to apply complexity analysis to nonlinear optimization problems, it is necessary to determine what it means to solve such a problem. Unlike linear problems, for nonlinear problems the length of the output can be infinite, such as in cases when a solution is irra- tional. There are two major complexity models for nonlinear optimization. One that seeks to approximate the objective function. The second, which we present here, approximates the optimal solution in the solution space. The latter has a number of advantages described next in Sect. 1.1. Our goals here are to set a framework for complexity analysis of nonlinear optimization with linear constraints, delineate as closely as possible the complexity borderlines between classes of nonlinear optimization problems, and generate a framework of effective tech- niques. In particular it is shown how properties of convexity, separability and quadraticness of nonlinear optimization contribute to a substantial reduction in the problems’ complexity. We review a spectrum of techniques that are most effective for each such class and demon- strate that in some cases the techniques lead to best possible algorithms in terms of their efficiency. This paper is an updated version of Hochbaum (2005). Some of the subjects covered appeared earlier in Hochbaum (1993).
1.1 The complexity model, or what constitutes a solution to a nonlinear optimization problem
There is a fundamental difficulty in solving nonlinear optimization on digital computers. Un- like linear programming, for which all basic solutions to a set of linear inequalities require only finite accuracy of polynomial length in the length of the input (see, e.g., Papadimitiou and Steiglitz 1982, Lemma 2.1), one cannot bound a priori the number of digits required for the length of nonlinear programming optimal solutions. This feature is important since computers can only store numbers of finite accuracy. Even the simplest nonlinear optimiza- tion problems can have irrational solutions and thus writing the output alone requires infi- nite complexity. This is the case, for instance, in the minimization of the convex function max{x2 − 2, 0}. So the interpretation of what it means to solve a problem in reals is not obvious. Traditional techniques for coping with nonlinear problems are reviewed extensively in Minoux (1986). These techniques have several shortcomings. In some applications, the non- linear function is not available analytically. It is accessible via a data acquisition process or as a solution to a system of differential equations (as in some queuing systems). A typical traditional method approximates first the data input function as an analytic function, while assuming it is a polynomial of a conveniently low degree. Consequently, there are errors attributed to the inaccuracy of the assumed input even before an algorithm is applied for solving the problem. Further difficulties arise because of conditions required for the algorithms to work. The algorithms typically make use of information about the function’s derivatives from numeri- cal approximations, which incorporates a further element of error in the eventual outcome. Moreover, certain assumptions about the properties of the nonlinear functions, such as dif- ferentiability and continuity of the gradients, are often made without any evidence that the “true” functions indeed possess them. Ann Oper Res (2007) 153: 257–296 259
In addition to the computational difficulties, the output for nonlinear continuous opti- mization problems can consist of irrational or even transcendental (non algebraic) numbers. Since there is no finite representation of such numbers, they are usually truncated to fit within the prescribed accuracy of the hardware and software. Complexity analysis requires finite length input and output which is of polynomial length as a function of the length of the input. Therefore the usual presentation of nonlinear prob- lems renders complexity analysis inapplicable. For this reason complexity theory has not addressed nonlinear problems, with the exception of some quadratic problems. To resolve this issue, the nonlinear functions can be given in form of a table or oracle. An oracle ac- cepts arguments of limited number of digits, and outputs the value of the function truncated to the prescribed number of digits. Since nonlinear functions cannot be treated with the same absolute accuracy as linear functions, the notion of approximate solutions is particularly important. One definition of a solution to a nonlinear optimization problem is one that approximates the objective func- tion. Among the major proponents of this approach are Nemirovsky and Yudin (1983)who chose to approximate the objective value while requiring some prior knowledge about the properties of the objective function. We observe that any information about the behavior of the objective at the optimum can always be translated to a level of accuracy of the solution vector itself (and vice versa). A detailed discussion of this point is provided in Hochbaum and Shanthikumar (1990). We thus believe that the interest of solving the optimization prob- lem is in terms of the accuracy of the solution rather than the accuracy of the optimal objec- tive value. We thus use a second definition of a solution is one that approximates the solution in the solution space, within a prescribed accuracy of . According to the concept of -accuracy of Hochbaum and Shanthikumar (1990)asolutionissaidtobe -accurate if it is at most ( ) at a distance of (in the L∞ norm) from an optimal solution. That is, a solution, x is ∗ ( ) ∗ -accurate if there exists an optimal solution x such that ||x − x ||∞ ≤ . is then said to be the accuracy required in the solution space. In other words, the solution is identical to 1 the optimum in O(log ) decimal digits. Both definitions of a solution overlap for nonlinear problems on integers in which case an accuracy of <1 is sufficient. The computation model here assumes the unit cost model, i.e. any arithmetic operation or comparison is counted as a single operation, even if the operands are real numbers. We 1 use here however only operands with up to O(log ) significant digits. A similar complexity model using -accuracy was used by Shub and Smale (1996); Rene- gar (1987) and others. These works, however, address algebraic problems, as opposed to optimization problems addressed here.
1.2 Classes of nonlinear problems addressed and some difficult cases
A fundamental concept linking between the solutions to continuous and integer problems is that of proximity. Classes of convex optimization problems for which there is a “good” proximity with relative closeness of the continuous and integer solutions are solvable in polynomial time that depends on that distance. These classes include problems with a con- straint matrix that has small subdeterminants. A prominent example of such optimization problems is the convex separable network flow problem. Other classes with a good proxim- ity addressed here are the problems with polymatroidal constraints (the allocation problem discussed in Sect. 5) and the NP-hard nonlinear knapsack problem. For the nonlinear knap- sack problem it is shown that approaches relying on proximity are effective in generating 260 Ann Oper Res (2007) 153: 257–296 a fully polynomial approximation scheme for the nonlinear knapsack problem as well as other results equivalent to those that hold for the linear knapsack problem. That is, the non- linearity of the objective function does not make the problem any harder. Among nonlinear problems with a constraint matrix that has very large subdeterminants, we discuss here the inverse shortest paths problem that has a constraint matrix the same as the dual of the multicommodity flow problem. For that problem we employ a proximity theorem called projected proximity that allows to get polynomial time algorithms for this class of nonlinear problems. Convex problems are discussed here with the classification: separable or nonseparable; quadratic or non-quadratic; integer or continuous. The nonconvex continuous version of the separable minimum cost network flow problem is NP-hard Sahni (1974)evenforthe quadratic case. The corresponding concave minimization problem is in general NP-hard (see, e.g., Guisewite and Pardalos 1990). An excellent unifying presentation of some polyno- mial instances of the concave separable cost minimization flow problem is given in Erickson et al. (1987). There the problem is proved to be polynomial when the arcs are incapacitated and number of demand nodes is fixed. The problem is also proved to have polynomial time algorithms for certain classes of planar graphs. The nonseparable nonlinear optimization problem is in general hard. Even with the as- sumption of convexity, a quadratic nonseparable problem is NP-hard (see Sect. 11.) In spite of these negative results, there are a number of subclasses of practical interest that are solv- able in polynomial time, which we point out. As a general rule, and as we show later, a convex separable nonquadratic integer flow problem is easier to solve than the respective continuous one, using a proximity result. The exception to this rule is the quadratic convex problems for which it is typically easier to obtain continuous solutions rather than integer ones. In that case an integer optimum is derived from the continuous solution instead of the other way around.
1.3 Polynomial algorithms and issues of strong polynomiality
All polynomial time algorithms presented here, and indeed all algorithms known, that solve nonlinear nonquadratic problems and run in polynomial time, do not have a strongly poly- nomial complexity. That is, the running time depends on the magnitude of some “number” in the data which is typically the range of the interval in which the variables are bounded. For example, the most efficient algorithms known to date for the convex integer separa- ble flow problem are based on the concept of scaling. These include an algorithm by Mi- noux (1986), an algorithm by Ahuja et al. (1993) and an algorithm of “proximity-scaling” type by Hochbaum and Shanthikumar (1990) (presented in Sect. 3). The complexity of B + + the proximity-scaling algorithm for the integer problem is O(log m (m n)(m n log n)) where B is the largest capacity or supply in the network and n, m, the number of nodes and arcs respectively. For the continuous problem the -accurate solution is derived in B + + O(log (m n)(m n log n)) steps. These three algorithms are polynomial but not strongly polynomial, as they depend, via the quantity B, on the magnitude of the numbers appearing in the problem instance, as well as on n and m. A naturally arising question is whether it is possible to devise an algorithm the complexity of which is independent of B and dependent only on n and m, that is, a strongly polynomial algorithm. This question was answered in the negative with an impossibility result for strongly polynomial algorithms in Hochbaum (1994) (Sect. 2). Strong polynomiality has emerged as an important issue since the first polynomial time algorithm, the ellipsoid method, was devised for solving linear programming problems. The Ann Oper Res (2007) 153: 257–296 261 ellipsoid method, as well as all other polynomial algorithms known for linear program- ming, runs polynomial but not strongly polynomial time. That is, the running time depends on the data coefficients rather than only on the number of variables and constraints. Con- sequently, solving linear programming problems (and other problems that do not possess strongly polynomial algorithms) with different degrees of accuracy in the cost coefficients results in different running times. So the actual number of arithmetic operations grows as the accuracy of the data, and hence the length of the numbers in the input, increases. Such behavior of an algorithm is undesirable as it requires careful monitoring of the size of the numbers appearing in the data describing the problem instance, and thus limits the efficient applicability of the algorithm. Although it is not known whether linear programming can be solved in strongly poly- nomial time, Tardos (1986) established that “combinatorial” linear programming problems, those with a constraint matrix having small coefficients, are solvable in strongly polyno- mial time. Thus, minimum (linear) cost network flow problems are solvable in strongly polynomial time (Tardos 1985) since their constraint matrix’ coefficients are either 0, 1 or −1. In contrast, nonlinear and non-quadratic optimization problems with linear constraints were proved impossible to solve in strongly polynomial time in a complexity model of the arithmetic operations, comparisons, and the rounding operation Hochbaum (1994). So while convex separable minimization is solved in polynomial time on totally unimodular matrices (Hochbaum and Shanthikumar 1990), linear optimization on such constraint matrices runs in strongly polynomial time. This negative result is not applicable to the quadratic case, and thus it may be possible to solve constrained quadratic optimization problems in strongly polynomial time. Yet, few quadratic optimization problems have been shown to be solvable in strongly polynomial time. For instance, it is not known how to solve the minimum quadratic cost network flow problem in strongly polynomial time. A number of special cases of the minimum quadratic cost network flow problem that are solvable in strong polynomial time are reviewed in Sect. 10.
1.4 Overview
We begin with the impossibility result and lower bound on the complexity of nonlinear prob- lems in Sect. 2. The lower bound provided applies in both the comparison model and in the algebraic tree model. Section 3 describes the proximity-scaling algorithm for convex sep- arable optimization problem with constraint matrices that have bounded subdeterminants. We focus on the interpretation of the technique as a form of piecewise linear approximation of nonlinear functions that uses specific scaling so as to guarantee polynomial complex- ity. A specific implementation of the approach to convex network flow is given in Sect. 4. In Sect. 5 we describe the use of a proximity-scaling approach to the general allocation problem and its special cases. Here the proximity theorem used is stronger than the one for general problems on linear constraints. The algorithms generated are shown to be best possi- ble for most classes of the allocation problem. The use of proximity to reduce the nonlinear knapsack problem to an allocation problem is described in Sect. 6. This leads to a poly- nomial time algorithm for the nonlinear continuous knapsack, which in turn makes a fully polynomial time approximation scheme available for the nonlinear knapsack problem. The use of proximity-scaling for the convex dual of minimum cost network flow is sketched next in Sect. 7. The proximity-scaling approach is concluded in Sect. 8 where a “projected prox- imity” theorem is shown to be applicable to the convex dual of the multi-commodity flow problem, with application to inverse shortest paths problem. 262 Ann Oper Res (2007) 153: 257–296
The next leading technique we describe is based on “threshold theorems” in Sect. 9.Itis shown there how this technique is used for the convex cost closure and the convex s-excess problem with one application, among many others, to the image segmentation problem. Classes of quadratic problems known to be solved in strongly polynomial time are de- lineated in Sect. 10. Various classes of nonseparable cases that are solved efficiently, and a collection of relevant techniques, are given in Sect. 11. Section 12 contains some conclud- ing remarks and lists open problems. Notation used in this paper includes bold letters for denoting vectors, and e to denote the vector (1, 1,...,1). The vector ej has 0s in all positions except position j that is equal to 1. When discussing a network the standard notation of G = (V, A) is used with the number of nodes |V | denoted by n and the number of arcs |A| denoted by m. In the discussion on non- linear programming the number of variables is denoted by n and the number of constraints (excluding nonnegativity) by m.
2 The impossibility of strongly polynomial algorithms for convex separable network flow problems
An impossibility result on the existence of a strongly polynomial algorithm for nonlinear problems was proved in Hochbaum (1994) and reviewed here. This result applies to con- vex separable minimization, and even to constraint matrices that are very simple, such as network flow constraints or even a single constraint bounding the sum of the values of the variables. This lower bound holds in the comparison model for all nonlinear problems. In a more general model—the algebraic-tree model that permits all the arithmetic operations— strongly polynomial algorithms are provably impossible for nonlinear and nonquadratic al- gorithms. That leaves open the possibility of strongly polynomial algorithms for quadratic convex separable minimization problems. The problem for which the lower bound is given is the simple resource allocation prob- lem. The simple resource allocation problem (denoted by the acronym SRA) is identical to a single source transportation problem in maximization form: n n (SRA) max fi (xi ) xi = B, x ≥ 0 . i=1 i=1 The generic presentation of this problem is as a concave maximization problem (with an obvious translation to the minimization/convex case). We first present a comparison model lower bound followed by an algebraic tree model lower bound.
2.1 A comparison model lower bound
A comparison computation model allows only the operations of comparisons and branch- ings. The lower bound proof establishes that no algorithm exists that solves SRA in less than log2 B comparisons. To show that, we rely on a result of information theory according to which there is no algorithm that finds a value in a monotone nonincreasing n-array that is the first to be smaller than some specified constant in less than log2 n time. Consider first a lower bound result for SRA in two variables, SRA(2):
(SRA(2)) max f1(x1) + cx2
x1 + x2 = B,
x1,x2 ≥ 0, integer. Ann Oper Res (2007) 153: 257–296 263
Let the function f1(x1) be given as an array of B increments at the B integer points. Namely, if f1(i) = ai the array of increments is {a0,a1 − a0,a2 − a1,...,aB − aB−1}.Since the function is concave, the entries in the array are monotone nonincreasing, ai+1 − ai ≤ ai − ai−1. The optimal solution to this problem is x1 = j and x2 = B − j,wherej is the largest index such that aj − aj−1 ≥ c. Since the array of the differences between consecutive entries of the array is monotone nonincreasing, determining in the array of differences the index j can be done using binary search in log2 B comparisons. The information theoretic lower bound is also log2 B compar- isons. This is because the comparison tree has B leaves so the path of comparisons leading from the root to a leaf could be as long as log2 B (see Knuth 1973). Suppose the problem could be solved independently of B, then given a monotone non- increasing array and a value c, it has a corresponding concave function, f1 such that the solution to the SRA(2) is independent of B. Consequently, the required entry in the vector could be found independently of B, which is a contradiction to the comparison tree lower bound. A similar proof works for the problem with a single variable if the constraint is an in- equality constraint. The same arguments can be extended to prove that in the comparison model the allocation + B problem on n 1 variables has complexity Ω(nlog2 n ). Let the problem be defined for c>0: n (SRA(n + 1)) max fj (xj ) + c · xn+1 j=1 n+1 xj = B, j=1
xj ≥ 0, integer, j = 1,...,n+ 1. [ B ] Let the functions fj be concave and monotone increasing in the interval 0, n , and zero [ B ] + B in n ,B . Solving SRA(n 1) is then equivalent to determining in n arrays of length n each, the last entry of value ≥ c. Since the arrays are independent, the information theory B lower bound is Ω(nlog n ). Similarly, for the case of an inequality constraint the same lower bound applies for the problem on n variables, since xn+1 can simply be viewed as the slack and c = 0. This comparison model lower bound holds also for the quadratic case. It is therefore impossible to solve the quadratic problems in strongly polynomial time using only compar- isons. Indeed the floor operation is essential for the quadratic integer problem and without it there is no hope of solving the integer version in strongly polynomial time This follows fromanobservationbyTamir(1993), that demonstrated this via the following quadratic allocation problem. 1 1 min x2 + (a − 1)x2 , 2 1 2 2
s.t.x1 + x2 = b,
x1,x2 ≥ 0, integer. B The optimal value of x2 is a . Therefore the floor operation can be executed via a rou- tine that solves a quadratic allocation problem. We demonstrate in the next section an impos- sibility result of strongly polynomial algorithms for non-quadratic problems which implies 264 Ann Oper Res (2007) 153: 257–296 that the floor operation, if helpful in devising strongly polynomial algorithms, is limited in its power to quadratic problems only.
2.2 The algebraic-tree model lower bound
One might criticize the choice of the comparison model as being too restrictive. Indeed, the use of arithmetic operations may help reduce the problems complexity. This is the case for the quadratic SRA, which is solvable in linear time, O(n) (Brucker 1984). The lower bound here demonstrates that such success is not possible for other nonlinear functions (non-quadratic). The algebraic-tree computation model permits the arithmetic operations +, −, ×, ÷ as well as comparisons and branchings based on any of these operations. It is demonstrated that the nature of the lower bound is unchanged even if the floor operation is permitted as well. We rely on Renegar’s lower bound proof (Renegar 1987) in this arithmetic model of com- putation for finding -accurate roots of polynomials of fixed degree ≥ 2. In particular, the [ ] R complexity of identifying an -accurate single real root in an interval O,R is Ω(log log ) even if the polynomial is monotone in that interval. Let p1(x), . . . , pn(x) be n polynomi- = [ B ] als each with a single root to the equation pi (x) c in the interval 0, n , and each pi (x) a monotone decreasing function in this interval. Since the choice of these polynomials is n n Ω(n B ) arbitrary, the lower bound on finding the roots of these polynomials is log log n . = xj ≥ Let fj (xj ) 0 pj (x) dx.Thefj s are then polynomials of degree 3. The problem, (P ) max fj (xj · ) + c · xn+1 · j + n 1 B x = , j j=1
xj ≥ 0integer,j= 1,...,n+ 1 has an optimal solution x such that y = · x is the (n )-accurate vector of roots solving the system ⎧ = ⎪ p1(y1) c, ⎨⎪ p2(y2) = c, ⎪ . ⎩⎪ . pn(yn) = c. This follows directly from the Kuhn–Tucker conditions of optimality and the proximity theorem to be discussed in Sect. 3.2, that an optimal integer solution x∗ to the scaled problem with a scaling constraint s and the optimal solution to the continuous problem y∗ satisfy ∗ ∗ x − y ∞ ≤ ns (Theorem 1). Hence, a lower bound for the complexity of solving (P ) is B = Ω(nlog log n ).For 1, we get the desired lower bound for the integer problem. Mansour et al. (1991) proved a lower bound on finding -accurate square roots that allows B also the floor, , operation. In our notation this lower bound is Ω( log log ). Hence, even with this additional operation the problem cannot be solved in strongly polynomial time. Again, the quadratic objective is an exception and indeed algorithms for solving the quadratic objective SRA problem rely on solving the continuous solution first, then rounding down, using the floor operation, and proceeding to compute the resulting integer vector to feasibility and optimality using fewer than n greedy steps. See for instance Ibaraki and Katoh Ann Oper Res (2007) 153: 257–296 265
(1988) for such an algorithm. Since the lower bound result applies also in the presence of the floor operation, it follows that the “ease” of solving the quadratic case is indeed due to the quadratic objective and not to this, perhaps powerful, operation.
3 A polynomial proximity-scaling algorithm
The proximity-scaling algorithm makes use of the fact that the convex piecewise linear min- imization problem on a “small” number of pieces is solvable in polynomial time (Dantzig 1963 and Lemma 1). The proximity-scaling algorithm is based on approximating the convex functions by piecewise linear functions on a uniform grid with a small number of break- points. The proximity theorems state that the solution to such scaled problems is close to the optimal solution in the solution space, thus allowing to update the length of the interval in which each variable lies by a constant factor. A logarithmic number of calls to solving the scaled piecewise linear function then leads to an optimal solution in integers, or within -accuracy. While the proximity-scaling procedure of Hochbaum and Shanthikumar (1990) provides a polynomial time algorithm for any convex separable optimization problem on totally uni- modular constraints (or for problems with constraint matrices that have bounded subdeter- minants), it is shown next that a specialized implementation taking into account the network structure and the equality balance constraints, is more efficient than the general purpose algorithm.
3.1 The scaled piecewise linear approximation
The idea of (piecewise) linearizing a nonlinear function in order to obtain solutions has been well known. In a 1959 book Dennis (1959) writes regarding quadratic cost networks: The electrical model for network flow problems can be extended to include flow branches for which the total cost contains terms depending on the square of the in- dividualbranchflows....Itappearsthatthealgorithmspresentedinthischaptercould begeneralized.... Thesemethods however are combinatorial in character and could require prohibitive calculation time, even on relatively simple networks. Certainly the simplicity and elegance of the diode-source algorithms would be absent. It would seem that the most practical means of attacking flow problems with quadratic costs would be to approximate the cost curve with piece-wise linear curve and substitute an appro- priate number of linear cost branches connected in parallel. Whereas it is clear that solving the problem on a piecewise linear approximation yields a feasible solution, the quality of such solution and its closeness to an optimal solution were not evaluated until the work of Hochbaum and Shanthikumar (1990). We consider a convex separable minimization problem
min{F(x | T x = b,≤ x ≤ u}.
The scaling process is illustrated for the convex network flow problem. For a network G = ={ } = (V, A) the variables are x xij (i,j)∈A and F(x) (i,j)∈A fij (xij ). s For a scaling constant s, the piecewise linearized objective function is F (x) = s ∈ s = (i,j)∈A fij (xij ) so that for each (i, j) Afij (xij ) fij (xij ) if xij is an integer mul- s tiple of s.Letfij (xij ) be defined so that it is linear in its argument between succes- 266 Ann Oper Res (2007) 153: 257–296
Fig. 1 Piecewise linear approximation
sive integral multiples of s. Thus each function f s (x ) is a piecewise linear approxi- ij ij s ij + = k p p = mation of fij as depicted in Fig. 1. Formally, fij (s s ks) p=1 Δij where Δij fij (ij + ps) − fij (ij + (p − 1)s). The convex integer network flow, (INF), problem is:
(INF) min F(x) s.t. T x = b, 0 ≤ x ≤ u, x, integer and its continuous (real) relaxation (RNF),
(RNF) min F(x) s.t. T x = b, 0 ≤ x ≤ u.
T is an n × m adjacency matrix of the network G = (V, A), b is a demand-supply n-vector, and u the capacity upper bounds vector on the arcs. When separable, the objective = function is F(x) (i,j)∈A fij (xij ). The (continuous) problem at the s-scaling phase, for any scaling constant s ∈ R+,isthe scaled problem (RNF-s) obtained by setting x = sy:
(RNF-s)minF s (sy) b s.t. T y = , s u 0 ≤ y ≤ . s Ann Oper Res (2007) 153: 257–296 267
Because of the equality constraints, the integer version (INF-s) (with the requirement that y is integer) has a feasible solution only if b/s is integer. Therefore, we choose an equivalent formulation of the scaled problem with inequality constraints. The scaled integer problem is then, (INF-s)minF s (sy) b s.t. T y ≥ , s b −T y ≥− , s u 0 ≤ y ≤ , s y integer. Although a feasible solution to (INF-s) is not necessarily feasible for the original prob- lem, yet the amount of unsatisfied supply (demand) is bounded by ( b −b ) · se = s s n bi −bi · ≤ 1 i=1( s s ) s ns units. The set of feasible solutions {x | T x = b, 0 ≤ x ≤ u} is bounded, a priori, in a box of length B = min{ u ∞, b 1} in each dimension as the flow on each edge cannot exceed − = uij ij capacity, or total sum of demands. Denoting Nij s , let each variable xij be substi- tuted by a sum of Nij variables each bounded between 0 and 1: N ij x = s ij + z(k) , 0 ≤ z(k) ≤ 1fork = 1,...,N. ij s ij ij k=1
In the network the analogous substitution is to replace each arc (i, j) by Nij arcs of capacity 1 each. The modified objective function for the linear programming formulation of both (LNF-s) and (INF-s) is thus,